Published 15 September 2026 · Primacy ITS

Use this to make a specific decision about a specific AI product, account type and task. Approval of a brand name alone is not enough: a personal account, business workspace, browser extension and connected agent can have different data flows and permissions.

The decision record

• Product and exact plan/workspace:
• Business task and expected benefit:
• Business owner and IT reviewer:
• Approved users or team:
• Allowed data categories:
• Prohibited data categories:
• Connections, plugins and permissions:
• Evidence links and date checked:
• Decision: not approved / limited pilot / approved for stated use:
• Conditions, expiry and next review:

1. Define the smallest useful task

Start with one real problem, such as rewriting a public help article or drafting a meeting agenda from non-confidential bullet points. Decide what a good result looks like and how a person will check it. Do not start by connecting every company file repository.

2. Trace what leaves the organisation

List prompts, uploads, recordings, transcripts, outputs and connected data sources. Record whether the tool can only read data or can also send messages, edit records, run code or make decisions. Assess additional access separately from the base chat tool.

3. Check the exact service terms

Use the vendor's documentation and contract for the actual plan. Record retention, deletion, model-training use, administrative controls, processors and data locations where relevant. If the answer is unclear, record it as unresolved and limit the pilot to public or synthetic information.

Do not treat a “we don't train on your data” statement as a complete privacy or security assessment. It does not by itself answer who can access the information, how long it is retained or what connected tools can do.

4. Give employees a short rule

Suggested wording for the approved-tool page:

“Use the listed business account for the listed tasks and data only. Check outputs before relying on them. Do not enter passwords, authentication codes, API keys or other secrets. Do not add connections or grant new permissions without approval.”

For tools not on the approved list:

“Do not put company-confidential, customer, employee or other personal information into unapproved AI tools or personal AI accounts. Ask IT for an approved way to do the task. Public or synthetic examples may be used only where the organisation's policy permits them.”

Approved business AI is not permission to upload everything. Sensitive use cases need their own assessment. Do not use the output to make consequential decisions about people without the appropriate human, legal and business review.

5. Run a limited pilot

Use a small group, low-risk inputs and a measurable task. Keep examples that demonstrate accuracy and failure modes without exposing confidential information. Check whether users can identify unsupported claims, accidental disclosure and instructions embedded in untrusted files.

6. Decide and publish the boundaries

Publish the exact tool/account, allowed task, allowed information, prohibited uses, reviewer and review date. Include a simple route to request another use case. If the pilot fails, revoke its connections and access and apply the agreed data-deletion process; preserve any required incident evidence first.

Verification

Ask a pilot user to show which account to use and classify three examples: a public product description, an employee appraisal and an API key. Check that the resulting choices match your written policy. Record the result and correct confusing wording before wider rollout.

This is a proposed operational checklist, not product-specific security assurance or a complete legal assessment. It deliberately makes no claims about a vendor's current plans. Verify each vendor's terms at the time of approval.

Reuse

Copy and adapt for internal use. Link this checklist to the organisation's acceptable-use policy, approved-tool register and incident reporting route. A full AI implementation kit can build on the record once tested with real users.

Related resources

Browse the toolkit library or suggest a question.